diff --git a/libcontainer/configs/validate/rootless.go b/libcontainer/configs/validate/rootless.go index 6d32704a32a..faceea04c09 100644 --- a/libcontainer/configs/validate/rootless.go +++ b/libcontainer/configs/validate/rootless.go @@ -2,6 +2,7 @@ package validate import ( "errors" + "fmt" "strconv" "strings" @@ -28,25 +29,18 @@ func rootlessEUIDCheck(config *configs.Config) error { return nil } -func hasIDMapping(id int, mappings []configs.IDMap) bool { - for _, m := range mappings { - if id >= m.ContainerID && id < m.ContainerID+m.Size { - return true - } - } - return false -} - func rootlessEUIDMappings(config *configs.Config) error { if !config.Namespaces.Contains(configs.NEWUSER) { return errors.New("rootless container requires user namespaces") } - - if len(config.UIDMappings) == 0 { - return errors.New("rootless containers requires at least one UID mapping") - } - if len(config.GIDMappings) == 0 { - return errors.New("rootless containers requires at least one GID mapping") + // We only require mappings if we are not joining another userns. + if path := config.Namespaces.PathOf(configs.NEWUSER); path == "" { + if len(config.UIDMappings) == 0 { + return errors.New("rootless containers requires at least one UID mapping") + } + if len(config.GIDMappings) == 0 { + return errors.New("rootless containers requires at least one GID mapping") + } } return nil } @@ -68,8 +62,8 @@ func rootlessEUIDMount(config *configs.Config) error { // Ignore unknown mount options. continue } - if !hasIDMapping(uid, config.UIDMappings) { - return errors.New("cannot specify uid= mount options for unmapped uid in rootless containers") + if _, err := config.HostUID(uid); err != nil { + return fmt.Errorf("cannot specify uid=%d mount option for rootless container: %w", uid, err) } } @@ -79,8 +73,8 @@ func rootlessEUIDMount(config *configs.Config) error { // Ignore unknown mount options. continue } - if !hasIDMapping(gid, config.GIDMappings) { - return errors.New("cannot specify gid= mount options for unmapped gid in rootless containers") + if _, err := config.HostGID(gid); err != nil { + return fmt.Errorf("cannot specify gid=%d mount option for rootless container: %w", gid, err) } } } diff --git a/libcontainer/specconv/spec_linux.go b/libcontainer/specconv/spec_linux.go index c1632ac12ad..2529e1e7e29 100644 --- a/libcontainer/specconv/spec_linux.go +++ b/libcontainer/specconv/spec_linux.go @@ -18,6 +18,7 @@ import ( "github.com/opencontainers/runc/libcontainer/configs" "github.com/opencontainers/runc/libcontainer/devices" "github.com/opencontainers/runc/libcontainer/seccomp" + "github.com/opencontainers/runc/libcontainer/userns" libcontainerUtils "github.com/opencontainers/runc/libcontainer/utils" "github.com/opencontainers/runtime-spec/specs-go" "github.com/sirupsen/logrus" @@ -968,6 +969,21 @@ func setupUserNamespace(spec *specs.Spec, config *configs.Config) error { config.UIDMappings = toConfigIDMap(spec.Linux.UIDMappings) config.GIDMappings = toConfigIDMap(spec.Linux.GIDMappings) } + if path := config.Namespaces.PathOf(configs.NEWUSER); path != "" { + // Cache the current userns mappings in our configuration, so that we + // can calculate uid and gid mappings within runc. These mappings are + // never used for configuring the container if the path is set. + uidMap, gidMap, err := userns.GetUserNamespaceMappings(path) + if err != nil { + return fmt.Errorf("failed to cache mappings for userns: %w", err) + } + config.UIDMappings = uidMap + config.GIDMappings = gidMap + logrus.WithFields(logrus.Fields{ + "uid_map": uidMap, + "gid_map": gidMap, + }).Debugf("config uses path-based userns configuration -- current uid and gid mappings cached") + } rootUID, err := config.HostRootUID() if err != nil { return err diff --git a/libcontainer/userns/userns_maps.c b/libcontainer/userns/userns_maps.c new file mode 100644 index 00000000000..84f2c6188c3 --- /dev/null +++ b/libcontainer/userns/userns_maps.c @@ -0,0 +1,79 @@ +#define _GNU_SOURCE +#include +#include +#include +#include +#include +#include + +/* + * All of the code here is run inside an aync-signal-safe context, so we need + * to be careful to not call any functions that could cause issues. In theory, + * since we are a Go program, there are fewer restrictions in practice, it's + * better to be safe than sorry. + * + * The only exception is exit, which we need to call to make sure we don't + * return into runc. + */ + +void bail(int pipefd, const char *fmt, ...) +{ + va_list args; + + va_start(args, fmt); + vdprintf(pipefd, fmt, args); + va_end(args); + + exit(1); +} + +int spawn_userns_cat(char *userns_path, char *path, int outfd, int errfd) +{ + char buffer[4096] = { 0 }; + + pid_t child = fork(); + if (child != 0) + return child; + /* in child */ + + /* Join the target userns. */ + int nsfd = open(userns_path, O_RDONLY); + if (nsfd < 0) + bail(errfd, "open userns path %s failed: %m", userns_path); + + int err = setns(nsfd, CLONE_NEWUSER); + if (err < 0) + bail(errfd, "setns %s failed: %m", userns_path); + + close(nsfd); + + /* Pipe the requested file contents. */ + int fd = open(path, O_RDONLY); + if (fd < 0) + bail(errfd, "open %s in userns %s failed: %m", path, userns_path); + + int nread, ntotal = 0; + while ((nread = read(fd, buffer, sizeof(buffer))) != 0) { + if (nread < 0) + bail(errfd, "read bytes from %s failed (after %d total bytes read): %m", path, ntotal); + ntotal += nread; + + int nwritten = 0; + while (nwritten < nread) { + int n = write(outfd, buffer, nread - nwritten); + if (n < 0) + bail(errfd, "write %d bytes from %s failed (after %d bytes written): %m", + nread - nwritten, path, nwritten); + nwritten += n; + } + if (nread != nwritten) + bail(errfd, "mismatch for bytes read and written: %d read != %d written", nread, nwritten); + } + + close(fd); + close(outfd); + close(errfd); + + /* We must exit here, otherwise we would return into a forked runc. */ + exit(0); +} diff --git a/libcontainer/userns/userns_maps_linux.go b/libcontainer/userns/userns_maps_linux.go new file mode 100644 index 00000000000..d81290de2cb --- /dev/null +++ b/libcontainer/userns/userns_maps_linux.go @@ -0,0 +1,171 @@ +//go:build linux + +package userns + +import ( + "bufio" + "bytes" + "fmt" + "io" + "os" + "unsafe" + + "github.com/opencontainers/runc/libcontainer/configs" + "github.com/sirupsen/logrus" +) + +/* +#include +extern int spawn_userns_cat(char *userns_path, char *path, int outfd, int errfd); +*/ +import "C" + +func parseIdmapData(data []byte) (ms []configs.IDMap, err error) { + scanner := bufio.NewScanner(bytes.NewReader(data)) + for scanner.Scan() { + var m configs.IDMap + line := scanner.Text() + if _, err := fmt.Sscanf(line, "%d %d %d", &m.ContainerID, &m.HostID, &m.Size); err != nil { + return nil, fmt.Errorf("parsing id map failed: invalid format in line %q: %w", line, err) + } + ms = append(ms, m) + } + if err := scanner.Err(); err != nil { + return nil, fmt.Errorf("parsing id map failed: %w", err) + } + return ms, nil +} + +// Do something equivalent to nsenter --user= cat , but more +// efficiently. Returns the contents of the requested file from within the user +// namespace. +func spawnUserNamespaceCat(nsPath string, path string) ([]byte, error) { + rdr, wtr, err := os.Pipe() + if err != nil { + return nil, fmt.Errorf("create pipe for userns spawn failed: %w", err) + } + defer rdr.Close() + defer wtr.Close() + + errRdr, errWtr, err := os.Pipe() + if err != nil { + return nil, fmt.Errorf("create error pipe for userns spawn failed: %w", err) + } + defer errRdr.Close() + defer errWtr.Close() + + cNsPath := C.CString(nsPath) + defer C.free(unsafe.Pointer(cNsPath)) + cPath := C.CString(path) + defer C.free(unsafe.Pointer(cPath)) + + childPid := C.spawn_userns_cat(cNsPath, cPath, C.int(wtr.Fd()), C.int(errWtr.Fd())) + + if childPid < 0 { + return nil, fmt.Errorf("failed to spawn fork for userns") + } else if childPid == 0 { + // this should never happen + panic("runc executing inside fork child -- unsafe state!") + } + + // We are in the parent -- close the write end of the pipe before reading. + wtr.Close() + output, err := io.ReadAll(rdr) + rdr.Close() + if err != nil { + return nil, fmt.Errorf("reading from userns spawn failed: %w", err) + } + + // Ditto for the error pipe. + errWtr.Close() + errOutput, err := io.ReadAll(errRdr) + errRdr.Close() + if err != nil { + return nil, fmt.Errorf("reading from userns spawn error pipe failed: %w", err) + } + errOutput = bytes.TrimSpace(errOutput) + + // Clean up the child. + child, err := os.FindProcess(int(childPid)) + if err != nil { + return nil, fmt.Errorf("could not find userns spawn process: %w", err) + } + state, err := child.Wait() + if err != nil { + return nil, fmt.Errorf("failed to wait for userns spawn process: %w", err) + } + if !state.Success() { + errStr := string(errOutput) + if errStr == "" { + errStr = fmt.Sprintf("unknown error (status code %d)", state.ExitCode()) + } + return nil, fmt.Errorf("userns spawn: %s", errStr) + } else if len(errOutput) > 0 { + // We can just ignore weird output in the error pipe if the process + // didn't bail(), but for completeness output for debugging. + logrus.Debugf("userns spawn succeeded but unexpected error message found: %s", string(errOutput)) + } + // The subprocess succeeded, return whatever it wrote to the pipe. + return output, nil +} + +func GetUserNamespaceMappings(nsPath string) (uidMap, gidMap []configs.IDMap, err error) { + var ( + pid int + extra rune + tryFastPath bool + ) + + // nsPath is usually of the form /proc//ns/user, which means that we + // already have a pid that is part of the user namespace and thus we can + // just use the pid to read from /proc//*id_map. + // + // Note that Sscanf doesn't consume the whole input, so we check for any + // trailing data with %c. That way, we can be sure the pattern matched + // /proc/$pid/ns/user _exactly_ iff n === 1. + if n, _ := fmt.Sscanf(nsPath, "/proc/%d/ns/user%c", &pid, &extra); n == 1 { + tryFastPath = pid > 0 + } + + for _, mapType := range []struct { + name string + idMap *[]configs.IDMap + }{ + {"uid_map", &uidMap}, + {"gid_map", &gidMap}, + } { + var mapData []byte + + if tryFastPath { + path := fmt.Sprintf("/proc/%d/%s", pid, mapType.name) + data, err := os.ReadFile(path) + if err != nil { + // Do not error out here -- we need to try the slow path if the + // fast path failed. + logrus.Debugf("failed to use fast path to read %s from userns %s (error: %s), falling back to slow userns-join path", mapType.name, nsPath, err) + } else { + mapData = data + } + } else { + logrus.Debugf("cannot use fast path to read %s from userns %s, falling back to slow userns-join path", mapType.name, nsPath) + } + + if mapData == nil { + // We have to actually join the namespace if we cannot take the + // fast path. The path is resolved with respect to the child + // process, so just use /proc/self. + data, err := spawnUserNamespaceCat(nsPath, "/proc/self/"+mapType.name) + if err != nil { + return nil, nil, err + } + mapData = data + } + idMap, err := parseIdmapData(mapData) + if err != nil { + return nil, nil, fmt.Errorf("failed to parse %s of userns %s: %w", mapType.name, nsPath, err) + } + *mapType.idMap = idMap + } + + return uidMap, gidMap, nil +} diff --git a/tests/integration/userns.bats b/tests/integration/userns.bats index d52444cacc5..947681c9b7c 100644 --- a/tests/integration/userns.bats +++ b/tests/integration/userns.bats @@ -15,15 +15,25 @@ function setup() { mkdir -p rootfs/{proc,sys,tmp} mkdir -p rootfs/tmp/mount-{1,2} + to_umount_list="$(mktemp "$BATS_RUN_TMPDIR/userns-mounts.XXXXXX")" if [ $EUID -eq 0 ]; then update_config ' .linux.namespaces += [{"type": "user"}] | .linux.uidMappings += [{"hostID": 100000, "containerID": 0, "size": 65534}] - | .linux.gidMappings += [{"hostID": 100000, "containerID": 0, "size": 65534}] ' + | .linux.gidMappings += [{"hostID": 200000, "containerID": 0, "size": 65534}] ' fi } function teardown() { teardown_bundle + + if [ -v to_umount_list ]; then + while read -r mount_path; do + umount -l "$mount_path" || : + rm -f "$mount_path" + done <"$to_umount_list" + rm -f "$to_umount_list" + unset to_umount_list + fi } @test "userns with simple mount" { @@ -83,3 +93,74 @@ function teardown() { runc exec test_busybox cat /sys/fs/cgroup/{pids,memory}/tasks [ "$status" -eq 0 ] } + +@test "userns join other container userns" { + # Create a detached container with the id-mapping we want. + update_config '.process.args = ["sleep", "infinity"]' + runc run -d --console-socket "$CONSOLE_SOCKET" target_userns + [ "$status" -eq 0 ] + + # Configure our container to attach to the first container's userns. + target_pid="$(__runc state target_userns | jq .pid)" + update_config '.linux.namespaces |= map(if .type == "user" then (.path = "/proc/'"$target_pid"'/ns/" + .type) else . end) + | del(.linux.uidMappings) + | del(.linux.gidMappings)' + runc run -d --console-socket "$CONSOLE_SOCKET" in_userns + [ "$status" -eq 0 ] + + runc exec in_userns cat /proc/self/uid_map + [ "$status" -eq 0 ] + if [ $EUID -eq 0 ]; then + grep -E '^\s+0\s+100000\s+65534$' <<<"$output" + else + grep -E '^\s+0\s+'$EUID'\s+1$' <<<"$output" + fi + + runc exec in_userns cat /proc/self/gid_map + [ "$status" -eq 0 ] + if [ $EUID -eq 0 ]; then + grep -E '^\s+0\s+200000\s+65534$' <<<"$output" + else + grep -E '^\s+0\s+'$EUID'\s+1$' <<<"$output" + fi +} + +@test "userns join other container userns [bind-mounted nsfd]" { + requires root + + # Create a detached container with the id-mapping we want. + update_config '.process.args = ["sleep", "infinity"]' + runc run -d --console-socket "$CONSOLE_SOCKET" target_userns + [ "$status" -eq 0 ] + + # Bind-mount the first containers userns nsfd to a different path, to + # exercise the non-fast-path (where runc has to join the userns to get the + # mappings). + target_pid="$(__runc state target_userns | jq .pid)" + userns_path=$(mktemp "$BATS_RUN_TMPDIR/userns.XXXXXX") + mount --bind "/proc/$target_pid/ns/user" "$userns_path" + echo "$userns_path" >>"$to_umount_list" + + # Configure our container to attach to the first container's userns. + update_config '.linux.namespaces |= map(if .type == "user" then (.path = "'"$userns_path"'") else . end) + | del(.linux.uidMappings) + | del(.linux.gidMappings)' + runc run -d --console-socket "$CONSOLE_SOCKET" in_userns + [ "$status" -eq 0 ] + + runc exec in_userns cat /proc/self/uid_map + [ "$status" -eq 0 ] + if [ $EUID -eq 0 ]; then + grep -E '^\s+0\s+100000\s+65534$' <<<"$output" + else + grep -E '^\s+0\s+'$EUID'\s+1$' <<<"$output" + fi + + runc exec in_userns cat /proc/self/gid_map + [ "$status" -eq 0 ] + if [ $EUID -eq 0 ]; then + grep -E '^\s+0\s+200000\s+65534$' <<<"$output" + else + grep -E '^\s+0\s+'$EUID'\s+1$' <<<"$output" + fi +}